it’s the faces, the character shapes, but really that yellow hue across the image. Perhaps someone with some color theory knowledge could explain why ChatGPT generates images like that.
Basically AI is started to poison itself, and AI has no concept of color grading, so errors are adding up. I don’t know if GLAZE and other poisoning methods made the matters worse, but wouldn’t be surprised.
For me it’s the scale and perspective that stood out first. Both people are the same size on-screen but the one on the right is also supposed to be closer so they’re actually huge, but they also have a tiny chair. Both chairs are also pointed away from the TV which is as big as a person, but they’re also somehow not facing each other so the closer person still has to turn around. The seat on the left would have to be pushed right up against the wall but they somehow managed to fit a lamp behind it too.
It just feels very strange as someone with first-hand experience of 3d space.
Afaik imgen ai starts with a salt and pepper noise image (randomised grey values, or maybe even randomised colour values). It then pushes and pulls the image in the direction of its learned nodes and weights.
So typically an imgen ai will generate images with a very even distribution of value (light and dark). It will have difficulty generating a picture like a Rembrandt painting with high light/dark contrasts. It probably can’t generate pictures with a lot of dark and one small bright spot. At least not out of the box.
I wonder if it’s a white balance thing, as in the setting you’d see on a camera or in a post processing tool.
For instance, consider that “soft” or “warm” light bulbs (say 3000K and below) are common in cozy indoor areas. They cast a much more yellow color of light compared with a daylight bulb or actual daylight, which will look very blue in comparison.
It’s like the model detected that the image was people in a living room and it applied a warm white balance to the whole picture because most images of a family in the living room have warm lighting globally.
But since it is a machine and apparently has not yet been explicitly taught that comics generally have bright colors and no strange tints, then it does not adjust accordingly.
I wonder if that is even giving it too much credit. Maybe it’s just the deterioration from all the iterations of garbage in, garbage out.
I read that it’s due to model collapse - the AIs are now trained on their own AI-generated content, like all the auto generated ghibli-style images with yellow tint
Not necessarily color theory, but archived comics from the 1970s and prior tend to be yellowed from degradation on their original prints (which were then scanned). It could also be the colors available to print artists at the time which were more muted compared to today.
i dont know how but it just FEELS like AI generated
I’m the same, I know it’s AI, but the only thing I can point to is the yellow staining. It’s just a vibe thing.
Piss filter.
There really is a style to it.
it’s the faces, the character shapes, but really that yellow hue across the image. Perhaps someone with some color theory knowledge could explain why ChatGPT generates images like that.
Basically AI is started to poison itself, and AI has no concept of color grading, so errors are adding up. I don’t know if GLAZE and other poisoning methods made the matters worse, but wouldn’t be surprised.
For me it’s the scale and perspective that stood out first. Both people are the same size on-screen but the one on the right is also supposed to be closer so they’re actually huge, but they also have a tiny chair. Both chairs are also pointed away from the TV which is as big as a person, but they’re also somehow not facing each other so the closer person still has to turn around. The seat on the left would have to be pushed right up against the wall but they somehow managed to fit a lamp behind it too.
It just feels very strange as someone with first-hand experience of 3d space.
#humblebrag
Afaik imgen ai starts with a salt and pepper noise image (randomised grey values, or maybe even randomised colour values). It then pushes and pulls the image in the direction of its learned nodes and weights.
So typically an imgen ai will generate images with a very even distribution of value (light and dark). It will have difficulty generating a picture like a Rembrandt painting with high light/dark contrasts. It probably can’t generate pictures with a lot of dark and one small bright spot. At least not out of the box.
I wonder if it’s a white balance thing, as in the setting you’d see on a camera or in a post processing tool.
For instance, consider that “soft” or “warm” light bulbs (say 3000K and below) are common in cozy indoor areas. They cast a much more yellow color of light compared with a daylight bulb or actual daylight, which will look very blue in comparison.
It’s like the model detected that the image was people in a living room and it applied a warm white balance to the whole picture because most images of a family in the living room have warm lighting globally.
But since it is a machine and apparently has not yet been explicitly taught that comics generally have bright colors and no strange tints, then it does not adjust accordingly.
I wonder if that is even giving it too much credit. Maybe it’s just the deterioration from all the iterations of garbage in, garbage out.
What your world isn’t piss coloured? How boring.
…There’s a breaking bad joke to be made somewhere…
I read that it’s due to model collapse - the AIs are now trained on their own AI-generated content, like all the auto generated ghibli-style images with yellow tint
Not necessarily color theory, but archived comics from the 1970s and prior tend to be yellowed from degradation on their original prints (which were then scanned). It could also be the colors available to print artists at the time which were more muted compared to today.
Although it really only happened recently. If you look at older generated images, they don’t have the yellow colouration.
Someday artists are going to struggle to replicate this “look” in order to make memes and jokes.
On the sides of our cave walls.